1 /*
   2  * Copyright (c) 2007, 2017, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.  Oracle designates this
   8  * particular file as subject to the "Classpath" exception as provided
   9  * by Oracle in the LICENSE file that accompanied this code.
  10  *
  11  * This code is distributed in the hope that it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  14  * version 2 for more details (a copy is included in the LICENSE file that
  15  * accompanied this code).
  16  *
  17  * You should have received a copy of the GNU General Public License version
  18  * 2 along with this work; if not, write to the Free Software Foundation,
  19  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  20  *
  21  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  22  * or visit www.oracle.com if you need additional information or have any
  23  * questions.
  24  */
  25 
  26 package sun.java2d.marlin;
  27 
  28 import sun.awt.geom.PathConsumer2D;
  29 import static sun.java2d.marlin.OffHeapArray.SIZE_INT;
  30 import jdk.internal.misc.Unsafe;
  31 
  32 final class Renderer implements PathConsumer2D, MarlinRenderer {
  33 
  34     static final boolean DISABLE_RENDER = false;
  35 
  36     static final boolean ENABLE_BLOCK_FLAGS = MarlinProperties.isUseTileFlags();
  37     static final boolean ENABLE_BLOCK_FLAGS_HEURISTICS = MarlinProperties.isUseTileFlagsWithHeuristics();
  38 
  39     private static final int ALL_BUT_LSB = 0xFFFFFFFE;
  40     private static final int ERR_STEP_MAX = 0x7FFFFFFF; // = 2^31 - 1
  41 
  42     private static final double POWER_2_TO_32 = 0x1.0p32d;
  43 
  44     // use float to make tosubpix methods faster (no int to float conversion)
  45     static final float SUBPIXEL_SCALE_X = (float) SUBPIXEL_POSITIONS_X;
  46     static final float SUBPIXEL_SCALE_Y = (float) SUBPIXEL_POSITIONS_Y;
  47     static final int SUBPIXEL_MASK_X = SUBPIXEL_POSITIONS_X - 1;
  48     static final int SUBPIXEL_MASK_Y = SUBPIXEL_POSITIONS_Y - 1;
  49 
  50     // number of subpixels corresponding to a tile line
  51     private static final int SUBPIXEL_TILE
  52         = TILE_H << SUBPIXEL_LG_POSITIONS_Y;
  53 
  54     // 2048 (pixelSize) pixels (height) x 8 subpixels = 64K
  55     static final int INITIAL_BUCKET_ARRAY
  56         = INITIAL_PIXEL_DIM * SUBPIXEL_POSITIONS_Y;
  57 
  58     // crossing capacity = edges count / 4 ~ 1024
  59     static final int INITIAL_CROSSING_COUNT = INITIAL_EDGES_COUNT >> 2;
  60 
  61     public static final int WIND_EVEN_ODD = 0;
  62     public static final int WIND_NON_ZERO = 1;
  63 
  64     // common to all types of input path segments.
  65     // OFFSET as bytes
  66     // only integer values:
  67     public static final long OFF_CURX_OR  = 0;
  68     public static final long OFF_ERROR    = OFF_CURX_OR  + SIZE_INT;
  69     public static final long OFF_BUMP_X   = OFF_ERROR    + SIZE_INT;
  70     public static final long OFF_BUMP_ERR = OFF_BUMP_X   + SIZE_INT;
  71     public static final long OFF_NEXT     = OFF_BUMP_ERR + SIZE_INT;
  72     public static final long OFF_YMAX     = OFF_NEXT     + SIZE_INT;
  73 
  74     // size of one edge in bytes
  75     public static final int SIZEOF_EDGE_BYTES = (int)(OFF_YMAX + SIZE_INT);
  76 
  77     // curve break into lines
  78     // cubic error in subpixels to decrement step
  79     private static final float CUB_DEC_ERR_SUBPIX
  80         = MarlinProperties.getCubicDecD2() * (NORM_SUBPIXELS / 8.0f); // 1 pixel
  81     // cubic error in subpixels to increment step
  82     private static final float CUB_INC_ERR_SUBPIX
  83         = MarlinProperties.getCubicIncD1() * (NORM_SUBPIXELS / 8.0f); // 0.4 pixel
  84 
  85     // TestNonAARasterization (JDK-8170879): cubics
  86     // bad paths (59294/100000 == 59,29%, 94335 bad pixels (avg = 1,59), 3966 warnings (avg = 0,07)
  87 
  88     // cubic bind length to decrement step
  89     public static final float CUB_DEC_BND
  90         = 8.0f * CUB_DEC_ERR_SUBPIX;
  91     // cubic bind length to increment step
  92     public static final float CUB_INC_BND
  93         = 8.0f * CUB_INC_ERR_SUBPIX;
  94 
  95     // cubic countlg
  96     public static final int CUB_COUNT_LG = 2;
  97     // cubic count = 2^countlg
  98     private static final int CUB_COUNT = 1 << CUB_COUNT_LG;
  99     // cubic count^2 = 4^countlg
 100     private static final int CUB_COUNT_2 = 1 << (2 * CUB_COUNT_LG);
 101     // cubic count^3 = 8^countlg
 102     private static final int CUB_COUNT_3 = 1 << (3 * CUB_COUNT_LG);
 103     // cubic dt = 1 / count
 104     private static final float CUB_INV_COUNT = 1.0f / CUB_COUNT;
 105     // cubic dt^2 = 1 / count^2 = 1 / 4^countlg
 106     private static final float CUB_INV_COUNT_2 = 1.0f / CUB_COUNT_2;
 107     // cubic dt^3 = 1 / count^3 = 1 / 8^countlg
 108     private static final float CUB_INV_COUNT_3 = 1.0f / CUB_COUNT_3;
 109 
 110     // quad break into lines
 111     // quadratic error in subpixels
 112     private static final float QUAD_DEC_ERR_SUBPIX
 113         = MarlinProperties.getQuadDecD2() * (NORM_SUBPIXELS / 8.0f); // 0.5 pixel
 114 
 115     // TestNonAARasterization (JDK-8170879): quads
 116     // bad paths (62916/100000 == 62,92%, 103818 bad pixels (avg = 1,65), 6514 warnings (avg = 0,10)
 117 
 118     // quadratic bind length to decrement step
 119     public static final float QUAD_DEC_BND
 120         = 8.0f * QUAD_DEC_ERR_SUBPIX;
 121 
 122 //////////////////////////////////////////////////////////////////////////////
 123 //  SCAN LINE
 124 //////////////////////////////////////////////////////////////////////////////
 125     // crossings ie subpixel edge x coordinates
 126     private int[] crossings;
 127     // auxiliary storage for crossings (merge sort)
 128     private int[] aux_crossings;
 129 
 130     // indices into the segment pointer lists. They indicate the "active"
 131     // sublist in the segment lists (the portion of the list that contains
 132     // all the segments that cross the next scan line).
 133     private int edgeCount;
 134     private int[] edgePtrs;
 135     // auxiliary storage for edge pointers (merge sort)
 136     private int[] aux_edgePtrs;
 137 
 138     // max used for both edgePtrs and crossings (stats only)
 139     private int activeEdgeMaxUsed;
 140 
 141     // crossings ref (dirty)
 142     private final IntArrayCache.Reference crossings_ref;
 143     // edgePtrs ref (dirty)
 144     private final IntArrayCache.Reference edgePtrs_ref;
 145     // merge sort initial arrays (large enough to satisfy most usages) (1024)
 146     // aux_crossings ref (dirty)
 147     private final IntArrayCache.Reference aux_crossings_ref;
 148     // aux_edgePtrs ref (dirty)
 149     private final IntArrayCache.Reference aux_edgePtrs_ref;
 150 
 151 //////////////////////////////////////////////////////////////////////////////
 152 //  EDGE LIST
 153 //////////////////////////////////////////////////////////////////////////////
 154     private int edgeMinY = Integer.MAX_VALUE;
 155     private int edgeMaxY = Integer.MIN_VALUE;
 156     private float edgeMinX = Float.POSITIVE_INFINITY;
 157     private float edgeMaxX = Float.NEGATIVE_INFINITY;
 158 
 159     // edges [ints] stored in off-heap memory
 160     private final OffHeapArray edges;
 161 
 162     private int[] edgeBuckets;
 163     private int[] edgeBucketCounts; // 2*newedges + (1 if pruning needed)
 164     // used range for edgeBuckets / edgeBucketCounts
 165     private int buckets_minY;
 166     private int buckets_maxY;
 167 
 168     // edgeBuckets ref (clean)
 169     private final IntArrayCache.Reference edgeBuckets_ref;
 170     // edgeBucketCounts ref (clean)
 171     private final IntArrayCache.Reference edgeBucketCounts_ref;
 172 
 173     // Flattens using adaptive forward differencing. This only carries out
 174     // one iteration of the AFD loop. All it does is update AFD variables (i.e.
 175     // X0, Y0, D*[X|Y], COUNT; not variables used for computing scanline crossings).
 176     private void quadBreakIntoLinesAndAdd(float x0, float y0,
 177                                           final Curve c,
 178                                           final float x2, final float y2)
 179     {
 180         int count = 1; // dt = 1 / count
 181 
 182         // maximum(ddX|Y) = norm(dbx, dby) * dt^2 (= 1)
 183         float maxDD = Math.abs(c.dbx) + Math.abs(c.dby);
 184 
 185         final float _DEC_BND = QUAD_DEC_BND;
 186 
 187         while (maxDD >= _DEC_BND) {
 188             // divide step by half:
 189             maxDD /= 4.0f; // error divided by 2^2 = 4
 190 
 191             count <<= 1;
 192             if (DO_STATS) {
 193                 rdrCtx.stats.stat_rdr_quadBreak_dec.add(count);
 194             }
 195         }
 196 
 197         int nL = 0; // line count
 198         if (count > 1) {
 199             final float icount = 1.0f / count; // dt
 200             final float icount2 = icount * icount; // dt^2
 201 
 202             final float ddx = c.dbx * icount2;
 203             final float ddy = c.dby * icount2;
 204             float dx = c.bx * icount2 + c.cx * icount;
 205             float dy = c.by * icount2 + c.cy * icount;
 206 
 207             float x1, y1;
 208 
 209             while (--count > 0) {
 210                 x1 = x0 + dx;
 211                 dx += ddx;
 212                 y1 = y0 + dy;
 213                 dy += ddy;
 214 
 215                 addLine(x0, y0, x1, y1);
 216 
 217                 if (DO_STATS) { nL++; }
 218                 x0 = x1;
 219                 y0 = y1;
 220             }
 221         }
 222         addLine(x0, y0, x2, y2);
 223 
 224         if (DO_STATS) {
 225             rdrCtx.stats.stat_rdr_quadBreak.add(nL + 1);
 226         }
 227     }
 228 
 229     // x0, y0 and x3,y3 are the endpoints of the curve. We could compute these
 230     // using c.xat(0),c.yat(0) and c.xat(1),c.yat(1), but this might introduce
 231     // numerical errors, and our callers already have the exact values.
 232     // Another alternative would be to pass all the control points, and call
 233     // c.set here, but then too many numbers are passed around.
 234     private void curveBreakIntoLinesAndAdd(float x0, float y0,
 235                                            final Curve c,
 236                                            final float x3, final float y3)
 237     {
 238         int count           = CUB_COUNT;
 239         final float icount  = CUB_INV_COUNT;   // dt
 240         final float icount2 = CUB_INV_COUNT_2; // dt^2
 241         final float icount3 = CUB_INV_COUNT_3; // dt^3
 242 
 243         // the dx and dy refer to forward differencing variables, not the last
 244         // coefficients of the "points" polynomial
 245         float dddx, dddy, ddx, ddy, dx, dy;
 246         dddx = 2.0f * c.dax * icount3;
 247         dddy = 2.0f * c.day * icount3;
 248         ddx = dddx + c.dbx * icount2;
 249         ddy = dddy + c.dby * icount2;
 250         dx = c.ax * icount3 + c.bx * icount2 + c.cx * icount;
 251         dy = c.ay * icount3 + c.by * icount2 + c.cy * icount;
 252 
 253         // we use x0, y0 to walk the line
 254         float x1 = x0, y1 = y0;
 255         int nL = 0; // line count
 256 
 257         final float _DEC_BND = CUB_DEC_BND;
 258         final float _INC_BND = CUB_INC_BND;
 259 
 260         while (count > 0) {
 261             // divide step by half:
 262             while (Math.abs(ddx) + Math.abs(ddy) >= _DEC_BND) {
 263                 dddx /= 8.0f;
 264                 dddy /= 8.0f;
 265                 ddx = ddx / 4.0f - dddx;
 266                 ddy = ddy / 4.0f - dddy;
 267                 dx = (dx - ddx) / 2.0f;
 268                 dy = (dy - ddy) / 2.0f;
 269 
 270                 count <<= 1;
 271                 if (DO_STATS) {
 272                     rdrCtx.stats.stat_rdr_curveBreak_dec.add(count);
 273                 }
 274             }
 275 
 276             // double step:
 277             // can only do this on even "count" values, because we must divide count by 2
 278             while (count % 2 == 0
 279                    && Math.abs(dx) + Math.abs(dy) <= _INC_BND)
 280             {
 281                 dx = 2.0f * dx + ddx;
 282                 dy = 2.0f * dy + ddy;
 283                 ddx = 4.0f * (ddx + dddx);
 284                 ddy = 4.0f * (ddy + dddy);
 285                 dddx *= 8.0f;
 286                 dddy *= 8.0f;
 287 
 288                 count >>= 1;
 289                 if (DO_STATS) {
 290                     rdrCtx.stats.stat_rdr_curveBreak_inc.add(count);
 291                 }
 292             }
 293             if (--count > 0) {
 294                 x1 += dx;
 295                 dx += ddx;
 296                 ddx += dddx;
 297                 y1 += dy;
 298                 dy += ddy;
 299                 ddy += dddy;
 300             } else {
 301                 x1 = x3;
 302                 y1 = y3;
 303             }
 304 
 305             addLine(x0, y0, x1, y1);
 306 
 307             if (DO_STATS) { nL++; }
 308             x0 = x1;
 309             y0 = y1;
 310         }
 311         if (DO_STATS) {
 312             rdrCtx.stats.stat_rdr_curveBreak.add(nL);
 313         }
 314     }
 315 
 316     private void addLine(float x1, float y1, float x2, float y2) {
 317         if (DO_MONITORS) {
 318             rdrCtx.stats.mon_rdr_addLine.start();
 319         }
 320         if (DO_STATS) {
 321             rdrCtx.stats.stat_rdr_addLine.add(1);
 322         }
 323         int or = 1; // orientation of the line. 1 if y increases, 0 otherwise.
 324         if (y2 < y1) {
 325             or = 0;
 326             float tmp = y2;
 327             y2 = y1;
 328             y1 = tmp;
 329             tmp = x2;
 330             x2 = x1;
 331             x1 = tmp;
 332         }
 333 
 334         // convert subpixel coordinates [float] into pixel positions [int]
 335 
 336         // The index of the pixel that holds the next HPC is at ceil(trueY - 0.5)
 337         // Since y1 and y2 are biased by -0.5 in tosubpixy(), this is simply
 338         // ceil(y1) or ceil(y2)
 339         // upper integer (inclusive)
 340         final int firstCrossing = FloatMath.max(FloatMath.ceil_int(y1), boundsMinY);
 341 
 342         // note: use boundsMaxY (last Y exclusive) to compute correct coverage
 343         // upper integer (exclusive)
 344         final int lastCrossing  = FloatMath.min(FloatMath.ceil_int(y2), boundsMaxY);
 345 
 346         /* skip horizontal lines in pixel space and clip edges
 347            out of y range [boundsMinY; boundsMaxY] */
 348         if (firstCrossing >= lastCrossing) {
 349             if (DO_MONITORS) {
 350                 rdrCtx.stats.mon_rdr_addLine.stop();
 351             }
 352             if (DO_STATS) {
 353                 rdrCtx.stats.stat_rdr_addLine_skip.add(1);
 354             }
 355             return;
 356         }
 357 
 358         // edge min/max X/Y are in subpixel space (half-open interval):
 359         // note: Use integer crossings to ensure consistent range within
 360         // edgeBuckets / edgeBucketCounts arrays in case of NaN values (int = 0)
 361         if (firstCrossing < edgeMinY) {
 362             edgeMinY = firstCrossing;
 363         }
 364         if (lastCrossing > edgeMaxY) {
 365             edgeMaxY = lastCrossing;
 366         }
 367 
 368         // Use double-precision for improved accuracy:
 369         final double x1d   = x1;
 370         final double y1d   = y1;
 371         final double slope = (x1d - x2) / (y1d - y2);
 372 
 373         if (slope >= 0.0d) { // <==> x1 < x2
 374             if (x1 < edgeMinX) {
 375                 edgeMinX = x1;
 376             }
 377             if (x2 > edgeMaxX) {
 378                 edgeMaxX = x2;
 379             }
 380         } else {
 381             if (x2 < edgeMinX) {
 382                 edgeMinX = x2;
 383             }
 384             if (x1 > edgeMaxX) {
 385                 edgeMaxX = x1;
 386             }
 387         }
 388 
 389         // local variables for performance:
 390         final int _SIZEOF_EDGE_BYTES = SIZEOF_EDGE_BYTES;
 391 
 392         final OffHeapArray _edges = edges;
 393 
 394         // get free pointer (ie length in bytes)
 395         final int edgePtr = _edges.used;
 396 
 397         // use substraction to avoid integer overflow:
 398         if (_edges.length - edgePtr < _SIZEOF_EDGE_BYTES) {
 399             // suppose _edges.length > _SIZEOF_EDGE_BYTES
 400             // so doubling size is enough to add needed bytes
 401             // note: throw IOOB if neededSize > 2Gb:
 402             final long edgeNewSize = ArrayCacheConst.getNewLargeSize(
 403                                         _edges.length,
 404                                         edgePtr + _SIZEOF_EDGE_BYTES);
 405 
 406             if (DO_STATS) {
 407                 rdrCtx.stats.stat_rdr_edges_resizes.add(edgeNewSize);
 408             }
 409             _edges.resize(edgeNewSize);
 410         }
 411 
 412 
 413         final Unsafe _unsafe = OffHeapArray.UNSAFE;
 414         final long SIZE_INT = 4L;
 415         long addr   = _edges.address + edgePtr;
 416 
 417         // The x value must be bumped up to its position at the next HPC we will evaluate.
 418         // "firstcrossing" is the (sub)pixel number where the next crossing occurs
 419         // thus, the actual coordinate of the next HPC is "firstcrossing + 0.5"
 420         // so the Y distance we cover is "firstcrossing + 0.5 - trueY".
 421         // Note that since y1 (and y2) are already biased by -0.5 in tosubpixy(), we have
 422         // y1 = trueY - 0.5
 423         // trueY = y1 + 0.5
 424         // firstcrossing + 0.5 - trueY = firstcrossing + 0.5 - (y1 + 0.5)
 425         //                             = firstcrossing - y1
 426         // The x coordinate at that HPC is then:
 427         // x1_intercept = x1 + (firstcrossing - y1) * slope
 428         // The next VPC is then given by:
 429         // VPC index = ceil(x1_intercept - 0.5), or alternately
 430         // VPC index = floor(x1_intercept - 0.5 + 1 - epsilon)
 431         // epsilon is hard to pin down in floating point, but easy in fixed point, so if
 432         // we convert to fixed point then these operations get easier:
 433         // long x1_fixed = x1_intercept * 2^32;  (fixed point 32.32 format)
 434         // curx = next VPC = fixed_floor(x1_fixed - 2^31 + 2^32 - 1)
 435         //                 = fixed_floor(x1_fixed + 2^31 - 1)
 436         //                 = fixed_floor(x1_fixed + 0x7FFFFFFF)
 437         // and error       = fixed_fract(x1_fixed + 0x7FFFFFFF)
 438         final double x1_intercept = x1d + (firstCrossing - y1d) * slope;
 439 
 440         // inlined scalb(x1_intercept, 32):
 441         final long x1_fixed_biased = ((long) (POWER_2_TO_32 * x1_intercept))
 442                                      + 0x7FFFFFFFL;
 443         // curx:
 444         // last bit corresponds to the orientation
 445         _unsafe.putInt(addr, (((int) (x1_fixed_biased >> 31L)) & ALL_BUT_LSB) | or);
 446         addr += SIZE_INT;
 447         _unsafe.putInt(addr,  ((int)  x1_fixed_biased) >>> 1);
 448         addr += SIZE_INT;
 449 
 450         // inlined scalb(slope, 32):
 451         final long slope_fixed = (long) (POWER_2_TO_32 * slope);
 452 
 453         // last bit set to 0 to keep orientation:
 454         _unsafe.putInt(addr, (((int) (slope_fixed >> 31L)) & ALL_BUT_LSB));
 455         addr += SIZE_INT;
 456         _unsafe.putInt(addr,  ((int)  slope_fixed) >>> 1);
 457         addr += SIZE_INT;
 458 
 459         final int[] _edgeBuckets      = edgeBuckets;
 460         final int[] _edgeBucketCounts = edgeBucketCounts;
 461 
 462         final int _boundsMinY = boundsMinY;
 463 
 464         // each bucket is a linked list. this method adds ptr to the
 465         // start of the "bucket"th linked list.
 466         final int bucketIdx = firstCrossing - _boundsMinY;
 467 
 468         // pointer from bucket
 469         _unsafe.putInt(addr, _edgeBuckets[bucketIdx]);
 470         addr += SIZE_INT;
 471         // y max (exclusive)
 472         _unsafe.putInt(addr,  lastCrossing);
 473 
 474         // Update buckets:
 475         // directly the edge struct "pointer"
 476         _edgeBuckets[bucketIdx]       = edgePtr;
 477         _edgeBucketCounts[bucketIdx] += 2; // 1 << 1
 478         // last bit means edge end
 479         _edgeBucketCounts[lastCrossing - _boundsMinY] |= 0x1;
 480 
 481         // update free pointer (ie length in bytes)
 482         _edges.used += _SIZEOF_EDGE_BYTES;
 483 
 484         if (DO_MONITORS) {
 485             rdrCtx.stats.mon_rdr_addLine.stop();
 486         }
 487     }
 488 
 489 // END EDGE LIST
 490 //////////////////////////////////////////////////////////////////////////////
 491 
 492     // Cache to store RLE-encoded coverage mask of the current primitive
 493     final MarlinCache cache;
 494 
 495     // Bounds of the drawing region, at subpixel precision.
 496     private int boundsMinX, boundsMinY, boundsMaxX, boundsMaxY;
 497 
 498     // Current winding rule
 499     private int windingRule;
 500 
 501     // Current drawing position, i.e., final point of last segment
 502     private float x0, y0;
 503 
 504     // Position of most recent 'moveTo' command
 505     private float sx0, sy0;
 506 
 507     // per-thread renderer context
 508     final RendererContext rdrCtx;
 509     // dirty curve
 510     private final Curve curve;
 511 
 512     // clean alpha array (zero filled)
 513     private int[] alphaLine;
 514 
 515     // alphaLine ref (clean)
 516     private final IntArrayCache.Reference alphaLine_ref;
 517 
 518     private boolean enableBlkFlags = false;
 519     private boolean prevUseBlkFlags = false;
 520 
 521     /* block flags (0|1) */
 522     private int[] blkFlags;
 523 
 524     // blkFlags ref (clean)
 525     private final IntArrayCache.Reference blkFlags_ref;
 526 
 527     Renderer(final RendererContext rdrCtx) {
 528         this.rdrCtx = rdrCtx;
 529 
 530         this.edges = rdrCtx.newOffHeapArray(INITIAL_EDGES_CAPACITY); // 96K
 531 
 532         this.curve = rdrCtx.curve;
 533 
 534         edgeBuckets_ref      = rdrCtx.newCleanIntArrayRef(INITIAL_BUCKET_ARRAY); // 64K
 535         edgeBucketCounts_ref = rdrCtx.newCleanIntArrayRef(INITIAL_BUCKET_ARRAY); // 64K
 536 
 537         edgeBuckets      = edgeBuckets_ref.initial;
 538         edgeBucketCounts = edgeBucketCounts_ref.initial;
 539 
 540         // 2048 (pixelsize) pixel large
 541         alphaLine_ref = rdrCtx.newCleanIntArrayRef(INITIAL_AA_ARRAY); // 8K
 542         alphaLine     = alphaLine_ref.initial;
 543 
 544         this.cache = rdrCtx.cache;
 545 
 546         crossings_ref     = rdrCtx.newDirtyIntArrayRef(INITIAL_CROSSING_COUNT); // 2K
 547         aux_crossings_ref = rdrCtx.newDirtyIntArrayRef(INITIAL_CROSSING_COUNT); // 2K
 548         edgePtrs_ref      = rdrCtx.newDirtyIntArrayRef(INITIAL_CROSSING_COUNT); // 2K
 549         aux_edgePtrs_ref  = rdrCtx.newDirtyIntArrayRef(INITIAL_CROSSING_COUNT); // 2K
 550 
 551         crossings     = crossings_ref.initial;
 552         aux_crossings = aux_crossings_ref.initial;
 553         edgePtrs      = edgePtrs_ref.initial;
 554         aux_edgePtrs  = aux_edgePtrs_ref.initial;
 555 
 556         blkFlags_ref = rdrCtx.newCleanIntArrayRef(INITIAL_ARRAY); // 1K = 1 tile line
 557         blkFlags     = blkFlags_ref.initial;
 558     }
 559 
 560     Renderer init(final int pix_boundsX, final int pix_boundsY,
 561                   final int pix_boundsWidth, final int pix_boundsHeight,
 562                   final int windingRule)
 563     {
 564         this.windingRule = windingRule;
 565 
 566         // bounds as half-open intervals: minX <= x < maxX and minY <= y < maxY
 567         this.boundsMinX =  pix_boundsX << SUBPIXEL_LG_POSITIONS_X;
 568         this.boundsMaxX =
 569             (pix_boundsX + pix_boundsWidth) << SUBPIXEL_LG_POSITIONS_X;
 570         this.boundsMinY =  pix_boundsY << SUBPIXEL_LG_POSITIONS_Y;
 571         this.boundsMaxY =
 572             (pix_boundsY + pix_boundsHeight) << SUBPIXEL_LG_POSITIONS_Y;
 573 
 574         if (DO_LOG_BOUNDS) {
 575             MarlinUtils.logInfo("boundsXY = [" + boundsMinX + " ... "
 576                                 + boundsMaxX + "[ [" + boundsMinY + " ... "
 577                                 + boundsMaxY + "[");
 578         }
 579 
 580         // see addLine: ceil(boundsMaxY) => boundsMaxY + 1
 581         // +1 for edgeBucketCounts
 582         final int edgeBucketsLength = (boundsMaxY - boundsMinY) + 1;
 583 
 584         if (edgeBucketsLength > INITIAL_BUCKET_ARRAY) {
 585             if (DO_STATS) {
 586                 rdrCtx.stats.stat_array_renderer_edgeBuckets
 587                     .add(edgeBucketsLength);
 588                 rdrCtx.stats.stat_array_renderer_edgeBucketCounts
 589                     .add(edgeBucketsLength);
 590             }
 591             edgeBuckets = edgeBuckets_ref.getArray(edgeBucketsLength);
 592             edgeBucketCounts = edgeBucketCounts_ref.getArray(edgeBucketsLength);
 593         }
 594 
 595         edgeMinY = Integer.MAX_VALUE;
 596         edgeMaxY = Integer.MIN_VALUE;
 597         edgeMinX = Float.POSITIVE_INFINITY;
 598         edgeMaxX = Float.NEGATIVE_INFINITY;
 599 
 600         // reset used mark:
 601         edgeCount = 0;
 602         activeEdgeMaxUsed = 0;
 603         edges.used = 0;
 604 
 605         return this; // fluent API
 606     }
 607 
 608     /**
 609      * Disposes this renderer and recycle it clean up before reusing this instance
 610      */
 611     void dispose() {
 612         if (DO_STATS) {
 613             rdrCtx.stats.stat_rdr_activeEdges.add(activeEdgeMaxUsed);
 614             rdrCtx.stats.stat_rdr_edges.add(edges.used);
 615             rdrCtx.stats.stat_rdr_edges_count.add(edges.used / SIZEOF_EDGE_BYTES);
 616             rdrCtx.stats.hist_rdr_edges_count.add(edges.used / SIZEOF_EDGE_BYTES);
 617             rdrCtx.stats.totalOffHeap += edges.length;
 618         }
 619         // Return arrays:
 620         crossings = crossings_ref.putArray(crossings);
 621         aux_crossings = aux_crossings_ref.putArray(aux_crossings);
 622 
 623         edgePtrs = edgePtrs_ref.putArray(edgePtrs);
 624         aux_edgePtrs = aux_edgePtrs_ref.putArray(aux_edgePtrs);
 625 
 626         alphaLine = alphaLine_ref.putArray(alphaLine, 0, 0); // already zero filled
 627         blkFlags  = blkFlags_ref.putArray(blkFlags, 0, 0); // already zero filled
 628 
 629         if (edgeMinY != Integer.MAX_VALUE) {
 630             // if context is maked as DIRTY:
 631             if (rdrCtx.dirty) {
 632                 // may happen if an exception if thrown in the pipeline processing:
 633                 // clear completely buckets arrays:
 634                 buckets_minY = 0;
 635                 buckets_maxY = boundsMaxY - boundsMinY;
 636             }
 637             // clear only used part
 638             edgeBuckets = edgeBuckets_ref.putArray(edgeBuckets, buckets_minY,
 639                                                                 buckets_maxY);
 640             edgeBucketCounts = edgeBucketCounts_ref.putArray(edgeBucketCounts,
 641                                                              buckets_minY,
 642                                                              buckets_maxY + 1);
 643         } else {
 644             // unused arrays
 645             edgeBuckets = edgeBuckets_ref.putArray(edgeBuckets, 0, 0);
 646             edgeBucketCounts = edgeBucketCounts_ref.putArray(edgeBucketCounts, 0, 0);
 647         }
 648 
 649         // At last: resize back off-heap edges to initial size
 650         if (edges.length != INITIAL_EDGES_CAPACITY) {
 651             // note: may throw OOME:
 652             edges.resize(INITIAL_EDGES_CAPACITY);
 653         }
 654         if (DO_CLEAN_DIRTY) {
 655             // Force zero-fill dirty arrays:
 656             edges.fill(BYTE_0);
 657         }
 658         if (DO_MONITORS) {
 659             rdrCtx.stats.mon_rdr_endRendering.stop();
 660         }
 661         // recycle the RendererContext instance
 662         MarlinRenderingEngine.returnRendererContext(rdrCtx);
 663     }
 664 
 665     private static float tosubpixx(final float pix_x) {
 666         return SUBPIXEL_SCALE_X * pix_x;
 667     }
 668 
 669     private static float tosubpixy(final float pix_y) {
 670         // shift y by -0.5 for fast ceil(y - 0.5):
 671         return SUBPIXEL_SCALE_Y * pix_y - 0.5f;
 672     }
 673 
 674     @Override
 675     public void moveTo(float pix_x0, float pix_y0) {
 676         closePath();
 677         final float sx = tosubpixx(pix_x0);
 678         final float sy = tosubpixy(pix_y0);
 679         this.sx0 = sx;
 680         this.sy0 = sy;
 681         this.x0 = sx;
 682         this.y0 = sy;
 683     }
 684 
 685     @Override
 686     public void lineTo(float pix_x1, float pix_y1) {
 687         final float x1 = tosubpixx(pix_x1);
 688         final float y1 = tosubpixy(pix_y1);
 689         addLine(x0, y0, x1, y1);
 690         x0 = x1;
 691         y0 = y1;
 692     }
 693 
 694     @Override
 695     public void curveTo(float x1, float y1,
 696                         float x2, float y2,
 697                         float x3, float y3)
 698     {
 699         final float xe = tosubpixx(x3);
 700         final float ye = tosubpixy(y3);
 701         curve.set(x0, y0, tosubpixx(x1), tosubpixy(y1),
 702                           tosubpixx(x2), tosubpixy(y2), xe, ye);
 703         curveBreakIntoLinesAndAdd(x0, y0, curve, xe, ye);
 704         x0 = xe;
 705         y0 = ye;
 706     }
 707 
 708     @Override
 709     public void quadTo(float x1, float y1, float x2, float y2) {
 710         final float xe = tosubpixx(x2);
 711         final float ye = tosubpixy(y2);
 712         curve.set(x0, y0, tosubpixx(x1), tosubpixy(y1), xe, ye);
 713         quadBreakIntoLinesAndAdd(x0, y0, curve, xe, ye);
 714         x0 = xe;
 715         y0 = ye;
 716     }
 717 
 718     @Override
 719     public void closePath() {
 720         addLine(x0, y0, sx0, sy0);
 721         x0 = sx0;
 722         y0 = sy0;
 723     }
 724 
 725     @Override
 726     public void pathDone() {
 727         closePath();
 728     }
 729 
 730     @Override
 731     public long getNativeConsumer() {
 732         throw new InternalError("Renderer does not use a native consumer.");
 733     }
 734 
 735     private void _endRendering(final int ymin, final int ymax) {
 736         if (DISABLE_RENDER) {
 737             return;
 738         }
 739 
 740         // Get X bounds as true pixel boundaries to compute correct pixel coverage:
 741         final int bboxx0 = bbox_spminX;
 742         final int bboxx1 = bbox_spmaxX;
 743 
 744         final boolean windingRuleEvenOdd = (windingRule == WIND_EVEN_ODD);
 745 
 746         // Useful when processing tile line by tile line
 747         final int[] _alpha = alphaLine;
 748 
 749         // local vars (performance):
 750         final MarlinCache _cache = cache;
 751         final OffHeapArray _edges = edges;
 752         final int[] _edgeBuckets = edgeBuckets;
 753         final int[] _edgeBucketCounts = edgeBucketCounts;
 754 
 755         int[] _crossings = this.crossings;
 756         int[] _edgePtrs  = this.edgePtrs;
 757 
 758         // merge sort auxiliary storage:
 759         int[] _aux_crossings = this.aux_crossings;
 760         int[] _aux_edgePtrs  = this.aux_edgePtrs;
 761 
 762         // copy constants:
 763         final long _OFF_ERROR    = OFF_ERROR;
 764         final long _OFF_BUMP_X   = OFF_BUMP_X;
 765         final long _OFF_BUMP_ERR = OFF_BUMP_ERR;
 766 
 767         final long _OFF_NEXT     = OFF_NEXT;
 768         final long _OFF_YMAX     = OFF_YMAX;
 769 
 770         final int _ALL_BUT_LSB   = ALL_BUT_LSB;
 771         final int _ERR_STEP_MAX  = ERR_STEP_MAX;
 772 
 773         // unsafe I/O:
 774         final Unsafe _unsafe = OffHeapArray.UNSAFE;
 775         final long    addr0  = _edges.address;
 776         long addr;
 777         final int _SUBPIXEL_LG_POSITIONS_X = SUBPIXEL_LG_POSITIONS_X;
 778         final int _SUBPIXEL_LG_POSITIONS_Y = SUBPIXEL_LG_POSITIONS_Y;
 779         final int _SUBPIXEL_MASK_X = SUBPIXEL_MASK_X;
 780         final int _SUBPIXEL_MASK_Y = SUBPIXEL_MASK_Y;
 781         final int _SUBPIXEL_POSITIONS_X = SUBPIXEL_POSITIONS_X;
 782 
 783         final int _MIN_VALUE = Integer.MIN_VALUE;
 784         final int _MAX_VALUE = Integer.MAX_VALUE;
 785 
 786         // Now we iterate through the scanlines. We must tell emitRow the coord
 787         // of the first non-transparent pixel, so we must keep accumulators for
 788         // the first and last pixels of the section of the current pixel row
 789         // that we will emit.
 790         // We also need to accumulate pix_bbox, but the iterator does it
 791         // for us. We will just get the values from it once this loop is done
 792         int minX = _MAX_VALUE;
 793         int maxX = _MIN_VALUE;
 794 
 795         int y = ymin;
 796         int bucket = y - boundsMinY;
 797 
 798         int numCrossings = this.edgeCount;
 799         int edgePtrsLen = _edgePtrs.length;
 800         int crossingsLen = _crossings.length;
 801         int _arrayMaxUsed = activeEdgeMaxUsed;
 802         int ptrLen = 0, newCount, ptrEnd;
 803 
 804         int bucketcount, i, j, ecur;
 805         int cross, lastCross;
 806         int x0, x1, tmp, sum, prev, curx, curxo, crorientation, err;
 807         int pix_x, pix_xmaxm1, pix_xmax;
 808 
 809         int low, high, mid, prevNumCrossings;
 810         boolean useBinarySearch;
 811 
 812         final int[] _blkFlags = blkFlags;
 813         final int _BLK_SIZE_LG = BLOCK_SIZE_LG;
 814         final int _BLK_SIZE = BLOCK_SIZE;
 815 
 816         final boolean _enableBlkFlagsHeuristics = ENABLE_BLOCK_FLAGS_HEURISTICS && this.enableBlkFlags;
 817 
 818         // Use block flags if large pixel span and few crossings:
 819         // ie mean(distance between crossings) is high
 820         boolean useBlkFlags = this.prevUseBlkFlags;
 821 
 822         final int stroking = rdrCtx.stroking;
 823 
 824         int lastY = -1; // last emited row
 825 
 826 
 827         // Iteration on scanlines
 828         for (; y < ymax; y++, bucket++) {
 829             // --- from former ScanLineIterator.next()
 830             bucketcount = _edgeBucketCounts[bucket];
 831 
 832             // marker on previously sorted edges:
 833             prevNumCrossings = numCrossings;
 834 
 835             // bucketCount indicates new edge / edge end:
 836             if (bucketcount != 0) {
 837                 if (DO_STATS) {
 838                     rdrCtx.stats.stat_rdr_activeEdges_updates.add(numCrossings);
 839                 }
 840 
 841                 // last bit set to 1 means that edges ends
 842                 if ((bucketcount & 0x1) != 0) {
 843                     // eviction in active edge list
 844                     // cache edges[] address + offset
 845                     addr = addr0 + _OFF_YMAX;
 846 
 847                     for (i = 0, newCount = 0; i < numCrossings; i++) {
 848                         // get the pointer to the edge
 849                         ecur = _edgePtrs[i];
 850                         // random access so use unsafe:
 851                         if (_unsafe.getInt(addr + ecur) > y) {
 852                             _edgePtrs[newCount++] = ecur;
 853                         }
 854                     }
 855                     // update marker on sorted edges minus removed edges:
 856                     prevNumCrossings = numCrossings = newCount;
 857                 }
 858 
 859                 ptrLen = bucketcount >> 1; // number of new edge
 860 
 861                 if (ptrLen != 0) {
 862                     if (DO_STATS) {
 863                         rdrCtx.stats.stat_rdr_activeEdges_adds.add(ptrLen);
 864                         if (ptrLen > 10) {
 865                             rdrCtx.stats.stat_rdr_activeEdges_adds_high.add(ptrLen);
 866                         }
 867                     }
 868                     ptrEnd = numCrossings + ptrLen;
 869 
 870                     if (edgePtrsLen < ptrEnd) {
 871                         if (DO_STATS) {
 872                             rdrCtx.stats.stat_array_renderer_edgePtrs.add(ptrEnd);
 873                         }
 874                         this.edgePtrs = _edgePtrs
 875                             = edgePtrs_ref.widenArray(_edgePtrs, numCrossings,
 876                                                       ptrEnd);
 877 
 878                         edgePtrsLen = _edgePtrs.length;
 879                         // Get larger auxiliary storage:
 880                         aux_edgePtrs_ref.putArray(_aux_edgePtrs);
 881 
 882                         // use ArrayCache.getNewSize() to use the same growing
 883                         // factor than widenArray():
 884                         if (DO_STATS) {
 885                             rdrCtx.stats.stat_array_renderer_aux_edgePtrs.add(ptrEnd);
 886                         }
 887                         this.aux_edgePtrs = _aux_edgePtrs
 888                             = aux_edgePtrs_ref.getArray(
 889                                 ArrayCacheConst.getNewSize(numCrossings, ptrEnd)
 890                             );
 891                     }
 892 
 893                     // cache edges[] address + offset
 894                     addr = addr0 + _OFF_NEXT;
 895 
 896                     // add new edges to active edge list:
 897                     for (ecur = _edgeBuckets[bucket];
 898                          numCrossings < ptrEnd; numCrossings++)
 899                     {
 900                         // store the pointer to the edge
 901                         _edgePtrs[numCrossings] = ecur;
 902                         // random access so use unsafe:
 903                         ecur = _unsafe.getInt(addr + ecur);
 904                     }
 905 
 906                     if (crossingsLen < numCrossings) {
 907                         // Get larger array:
 908                         crossings_ref.putArray(_crossings);
 909 
 910                         if (DO_STATS) {
 911                             rdrCtx.stats.stat_array_renderer_crossings
 912                                 .add(numCrossings);
 913                         }
 914                         this.crossings = _crossings
 915                             = crossings_ref.getArray(numCrossings);
 916 
 917                         // Get larger auxiliary storage:
 918                         aux_crossings_ref.putArray(_aux_crossings);
 919 
 920                         if (DO_STATS) {
 921                             rdrCtx.stats.stat_array_renderer_aux_crossings
 922                                 .add(numCrossings);
 923                         }
 924                         this.aux_crossings = _aux_crossings
 925                             = aux_crossings_ref.getArray(numCrossings);
 926 
 927                         crossingsLen = _crossings.length;
 928                     }
 929                     if (DO_STATS) {
 930                         // update max used mark
 931                         if (numCrossings > _arrayMaxUsed) {
 932                             _arrayMaxUsed = numCrossings;
 933                         }
 934                     }
 935                 } // ptrLen != 0
 936             } // bucketCount != 0
 937 
 938 
 939             if (numCrossings != 0) {
 940                 /*
 941                  * thresholds to switch to optimized merge sort
 942                  * for newly added edges + final merge pass.
 943                  */
 944                 if ((ptrLen < 10) || (numCrossings < 40)) {
 945                     if (DO_STATS) {
 946                         rdrCtx.stats.hist_rdr_crossings.add(numCrossings);
 947                         rdrCtx.stats.hist_rdr_crossings_adds.add(ptrLen);
 948                     }
 949 
 950                     /*
 951                      * threshold to use binary insertion sort instead of
 952                      * straight insertion sort (to reduce minimize comparisons).
 953                      */
 954                     useBinarySearch = (numCrossings >= 20);
 955 
 956                     // if small enough:
 957                     lastCross = _MIN_VALUE;
 958 
 959                     for (i = 0; i < numCrossings; i++) {
 960                         // get the pointer to the edge
 961                         ecur = _edgePtrs[i];
 962 
 963                         /* convert subpixel coordinates into pixel
 964                             positions for coming scanline */
 965                         /* note: it is faster to always update edges even
 966                            if it is removed from AEL for coming or last scanline */
 967 
 968                         // random access so use unsafe:
 969                         addr = addr0 + ecur; // ecur + OFF_F_CURX
 970 
 971                         // get current crossing:
 972                         curx = _unsafe.getInt(addr);
 973 
 974                         // update crossing with orientation at last bit:
 975                         cross = curx;
 976 
 977                         // Increment x using DDA (fixed point):
 978                         curx += _unsafe.getInt(addr + _OFF_BUMP_X);
 979 
 980                         // Increment error:
 981                         err  =  _unsafe.getInt(addr + _OFF_ERROR)
 982                               + _unsafe.getInt(addr + _OFF_BUMP_ERR);
 983 
 984                         // Manual carry handling:
 985                         // keep sign and carry bit only and ignore last bit (preserve orientation):
 986                         _unsafe.putInt(addr,               curx - ((err >> 30) & _ALL_BUT_LSB));
 987                         _unsafe.putInt(addr + _OFF_ERROR, (err & _ERR_STEP_MAX));
 988 
 989                         if (DO_STATS) {
 990                             rdrCtx.stats.stat_rdr_crossings_updates.add(numCrossings);
 991                         }
 992 
 993                         // insertion sort of crossings:
 994                         if (cross < lastCross) {
 995                             if (DO_STATS) {
 996                                 rdrCtx.stats.stat_rdr_crossings_sorts.add(i);
 997                             }
 998 
 999                             /* use binary search for newly added edges
1000                                in crossings if arrays are large enough */
1001                             if (useBinarySearch && (i >= prevNumCrossings)) {
1002                                 if (DO_STATS) {
1003                                     rdrCtx.stats.stat_rdr_crossings_bsearch.add(i);
1004                                 }
1005                                 low = 0;
1006                                 high = i - 1;
1007 
1008                                 do {
1009                                     // note: use signed shift (not >>>) for performance
1010                                     // as indices are small enough to exceed Integer.MAX_VALUE
1011                                     mid = (low + high) >> 1;
1012 
1013                                     if (_crossings[mid] < cross) {
1014                                         low = mid + 1;
1015                                     } else {
1016                                         high = mid - 1;
1017                                     }
1018                                 } while (low <= high);
1019 
1020                                 for (j = i - 1; j >= low; j--) {
1021                                     _crossings[j + 1] = _crossings[j];
1022                                     _edgePtrs [j + 1] = _edgePtrs[j];
1023                                 }
1024                                 _crossings[low] = cross;
1025                                 _edgePtrs [low] = ecur;
1026 
1027                             } else {
1028                                 j = i - 1;
1029                                 _crossings[i] = _crossings[j];
1030                                 _edgePtrs[i] = _edgePtrs[j];
1031 
1032                                 while ((--j >= 0) && (_crossings[j] > cross)) {
1033                                     _crossings[j + 1] = _crossings[j];
1034                                     _edgePtrs [j + 1] = _edgePtrs[j];
1035                                 }
1036                                 _crossings[j + 1] = cross;
1037                                 _edgePtrs [j + 1] = ecur;
1038                             }
1039 
1040                         } else {
1041                             _crossings[i] = lastCross = cross;
1042                         }
1043                     }
1044                 } else {
1045                     if (DO_STATS) {
1046                         rdrCtx.stats.stat_rdr_crossings_msorts.add(numCrossings);
1047                         rdrCtx.stats.hist_rdr_crossings_ratio
1048                             .add((1000 * ptrLen) / numCrossings);
1049                         rdrCtx.stats.hist_rdr_crossings_msorts.add(numCrossings);
1050                         rdrCtx.stats.hist_rdr_crossings_msorts_adds.add(ptrLen);
1051                     }
1052 
1053                     // Copy sorted data in auxiliary arrays
1054                     // and perform insertion sort on almost sorted data
1055                     // (ie i < prevNumCrossings):
1056 
1057                     lastCross = _MIN_VALUE;
1058 
1059                     for (i = 0; i < numCrossings; i++) {
1060                         // get the pointer to the edge
1061                         ecur = _edgePtrs[i];
1062 
1063                         /* convert subpixel coordinates into pixel
1064                             positions for coming scanline */
1065                         /* note: it is faster to always update edges even
1066                            if it is removed from AEL for coming or last scanline */
1067 
1068                         // random access so use unsafe:
1069                         addr = addr0 + ecur; // ecur + OFF_F_CURX
1070 
1071                         // get current crossing:
1072                         curx = _unsafe.getInt(addr);
1073 
1074                         // update crossing with orientation at last bit:
1075                         cross = curx;
1076 
1077                         // Increment x using DDA (fixed point):
1078                         curx += _unsafe.getInt(addr + _OFF_BUMP_X);
1079 
1080                         // Increment error:
1081                         err  =  _unsafe.getInt(addr + _OFF_ERROR)
1082                               + _unsafe.getInt(addr + _OFF_BUMP_ERR);
1083 
1084                         // Manual carry handling:
1085                         // keep sign and carry bit only and ignore last bit (preserve orientation):
1086                         _unsafe.putInt(addr,               curx - ((err >> 30) & _ALL_BUT_LSB));
1087                         _unsafe.putInt(addr + _OFF_ERROR, (err & _ERR_STEP_MAX));
1088 
1089                         if (DO_STATS) {
1090                             rdrCtx.stats.stat_rdr_crossings_updates.add(numCrossings);
1091                         }
1092 
1093                         if (i >= prevNumCrossings) {
1094                             // simply store crossing as edgePtrs is in-place:
1095                             // will be copied and sorted efficiently by mergesort later:
1096                             _crossings[i]     = cross;
1097 
1098                         } else if (cross < lastCross) {
1099                             if (DO_STATS) {
1100                                 rdrCtx.stats.stat_rdr_crossings_sorts.add(i);
1101                             }
1102 
1103                             // (straight) insertion sort of crossings:
1104                             j = i - 1;
1105                             _aux_crossings[i] = _aux_crossings[j];
1106                             _aux_edgePtrs[i] = _aux_edgePtrs[j];
1107 
1108                             while ((--j >= 0) && (_aux_crossings[j] > cross)) {
1109                                 _aux_crossings[j + 1] = _aux_crossings[j];
1110                                 _aux_edgePtrs [j + 1] = _aux_edgePtrs[j];
1111                             }
1112                             _aux_crossings[j + 1] = cross;
1113                             _aux_edgePtrs [j + 1] = ecur;
1114 
1115                         } else {
1116                             // auxiliary storage:
1117                             _aux_crossings[i] = lastCross = cross;
1118                             _aux_edgePtrs [i] = ecur;
1119                         }
1120                     }
1121 
1122                     // use Mergesort using auxiliary arrays (sort only right part)
1123                     MergeSort.mergeSortNoCopy(_crossings,     _edgePtrs,
1124                                               _aux_crossings, _aux_edgePtrs,
1125                                               numCrossings,   prevNumCrossings);
1126                 }
1127 
1128                 // reset ptrLen
1129                 ptrLen = 0;
1130                 // --- from former ScanLineIterator.next()
1131 
1132 
1133                 /* note: bboxx0 and bboxx1 must be pixel boundaries
1134                    to have correct coverage computation */
1135 
1136                 // right shift on crossings to get the x-coordinate:
1137                 curxo = _crossings[0];
1138                 x0    = curxo >> 1;
1139                 if (x0 < minX) {
1140                     minX = x0; // subpixel coordinate
1141                 }
1142 
1143                 x1 = _crossings[numCrossings - 1] >> 1;
1144                 if (x1 > maxX) {
1145                     maxX = x1; // subpixel coordinate
1146                 }
1147 
1148 
1149                 // compute pixel coverages
1150                 prev = curx = x0;
1151                 // to turn {0, 1} into {-1, 1}, multiply by 2 and subtract 1.
1152                 // last bit contains orientation (0 or 1)
1153                 crorientation = ((curxo & 0x1) << 1) - 1;
1154 
1155                 if (windingRuleEvenOdd) {
1156                     sum = crorientation;
1157 
1158                     // Even Odd winding rule: take care of mask ie sum(orientations)
1159                     for (i = 1; i < numCrossings; i++) {
1160                         curxo = _crossings[i];
1161                         curx  =  curxo >> 1;
1162                         // to turn {0, 1} into {-1, 1}, multiply by 2 and subtract 1.
1163                         // last bit contains orientation (0 or 1)
1164                         crorientation = ((curxo & 0x1) << 1) - 1;
1165 
1166                         if ((sum & 0x1) != 0) {
1167                             // TODO: perform line clipping on left-right sides
1168                             // to avoid such bound checks:
1169                             x0 = (prev > bboxx0) ? prev : bboxx0;
1170 
1171                             if (curx < bboxx1) {
1172                                 x1 = curx;
1173                             } else {
1174                                 x1 = bboxx1;
1175                                 // skip right side (fast exit loop):
1176                                 i = numCrossings;
1177                             }
1178 
1179                             if (x0 < x1) {
1180                                 x0 -= bboxx0; // turn x0, x1 from coords to indices
1181                                 x1 -= bboxx0; // in the alpha array.
1182 
1183                                 pix_x      =  x0      >> _SUBPIXEL_LG_POSITIONS_X;
1184                                 pix_xmaxm1 = (x1 - 1) >> _SUBPIXEL_LG_POSITIONS_X;
1185 
1186                                 if (pix_x == pix_xmaxm1) {
1187                                     // Start and end in same pixel
1188                                     tmp = (x1 - x0); // number of subpixels
1189                                     _alpha[pix_x    ] += tmp;
1190                                     _alpha[pix_x + 1] -= tmp;
1191 
1192                                     if (useBlkFlags) {
1193                                         // flag used blocks:
1194                                         // note: block processing handles extra pixel:
1195                                         _blkFlags[pix_x    >> _BLK_SIZE_LG] = 1;
1196                                     }
1197                                 } else {
1198                                     tmp = (x0 & _SUBPIXEL_MASK_X);
1199                                     _alpha[pix_x    ]
1200                                         += (_SUBPIXEL_POSITIONS_X - tmp);
1201                                     _alpha[pix_x + 1]
1202                                         += tmp;
1203 
1204                                     pix_xmax = x1 >> _SUBPIXEL_LG_POSITIONS_X;
1205 
1206                                     tmp = (x1 & _SUBPIXEL_MASK_X);
1207                                     _alpha[pix_xmax    ]
1208                                         -= (_SUBPIXEL_POSITIONS_X - tmp);
1209                                     _alpha[pix_xmax + 1]
1210                                         -= tmp;
1211 
1212                                     if (useBlkFlags) {
1213                                         // flag used blocks:
1214                                         // note: block processing handles extra pixel:
1215                                         _blkFlags[pix_x    >> _BLK_SIZE_LG] = 1;
1216                                         _blkFlags[pix_xmax >> _BLK_SIZE_LG] = 1;
1217                                     }
1218                                 }
1219                             }
1220                         }
1221 
1222                         sum += crorientation;
1223                         prev = curx;
1224                     }
1225                 } else {
1226                     // Non-zero winding rule: optimize that case (default)
1227                     // and avoid processing intermediate crossings
1228                     for (i = 1, sum = 0;; i++) {
1229                         sum += crorientation;
1230 
1231                         if (sum != 0) {
1232                             // prev = min(curx)
1233                             if (prev > curx) {
1234                                 prev = curx;
1235                             }
1236                         } else {
1237                             // TODO: perform line clipping on left-right sides
1238                             // to avoid such bound checks:
1239                             x0 = (prev > bboxx0) ? prev : bboxx0;
1240 
1241                             if (curx < bboxx1) {
1242                                 x1 = curx;
1243                             } else {
1244                                 x1 = bboxx1;
1245                                 // skip right side (fast exit loop):
1246                                 i = numCrossings;
1247                             }
1248 
1249                             if (x0 < x1) {
1250                                 x0 -= bboxx0; // turn x0, x1 from coords to indices
1251                                 x1 -= bboxx0; // in the alpha array.
1252 
1253                                 pix_x      =  x0      >> _SUBPIXEL_LG_POSITIONS_X;
1254                                 pix_xmaxm1 = (x1 - 1) >> _SUBPIXEL_LG_POSITIONS_X;
1255 
1256                                 if (pix_x == pix_xmaxm1) {
1257                                     // Start and end in same pixel
1258                                     tmp = (x1 - x0); // number of subpixels
1259                                     _alpha[pix_x    ] += tmp;
1260                                     _alpha[pix_x + 1] -= tmp;
1261 
1262                                     if (useBlkFlags) {
1263                                         // flag used blocks:
1264                                         // note: block processing handles extra pixel:
1265                                         _blkFlags[pix_x    >> _BLK_SIZE_LG] = 1;
1266                                     }
1267                                 } else {
1268                                     tmp = (x0 & _SUBPIXEL_MASK_X);
1269                                     _alpha[pix_x    ]
1270                                         += (_SUBPIXEL_POSITIONS_X - tmp);
1271                                     _alpha[pix_x + 1]
1272                                         += tmp;
1273 
1274                                     pix_xmax = x1 >> _SUBPIXEL_LG_POSITIONS_X;
1275 
1276                                     tmp = (x1 & _SUBPIXEL_MASK_X);
1277                                     _alpha[pix_xmax    ]
1278                                         -= (_SUBPIXEL_POSITIONS_X - tmp);
1279                                     _alpha[pix_xmax + 1]
1280                                         -= tmp;
1281 
1282                                     if (useBlkFlags) {
1283                                         // flag used blocks:
1284                                         // note: block processing handles extra pixel:
1285                                         _blkFlags[pix_x    >> _BLK_SIZE_LG] = 1;
1286                                         _blkFlags[pix_xmax >> _BLK_SIZE_LG] = 1;
1287                                     }
1288                                 }
1289                             }
1290                             prev = _MAX_VALUE;
1291                         }
1292 
1293                         if (i == numCrossings) {
1294                             break;
1295                         }
1296 
1297                         curxo = _crossings[i];
1298                         curx  =  curxo >> 1;
1299                         // to turn {0, 1} into {-1, 1}, multiply by 2 and subtract 1.
1300                         // last bit contains orientation (0 or 1)
1301                         crorientation = ((curxo & 0x1) << 1) - 1;
1302                     }
1303                 }
1304             } // numCrossings > 0
1305 
1306             // even if this last row had no crossings, alpha will be zeroed
1307             // from the last emitRow call. But this doesn't matter because
1308             // maxX < minX, so no row will be emitted to the MarlinCache.
1309             if ((y & _SUBPIXEL_MASK_Y) == _SUBPIXEL_MASK_Y) {
1310                 lastY = y >> _SUBPIXEL_LG_POSITIONS_Y;
1311 
1312                 // convert subpixel to pixel coordinate within boundaries:
1313                 minX = FloatMath.max(minX, bboxx0) >> _SUBPIXEL_LG_POSITIONS_X;
1314                 maxX = FloatMath.min(maxX, bboxx1) >> _SUBPIXEL_LG_POSITIONS_X;
1315 
1316                 if (maxX >= minX) {
1317                     // note: alpha array will be zeroed by copyAARow()
1318                     // +1 because alpha [pix_minX; pix_maxX[
1319                     // fix range [x0; x1[
1320                     // note: if x1=bboxx1, then alpha is written up to bboxx1+1
1321                     // inclusive: alpha[bboxx1] ignored, alpha[bboxx1+1] == 0
1322                     // (normally so never cleared below)
1323                     copyAARow(_alpha, lastY, minX, maxX + 1, useBlkFlags);
1324 
1325                     // speculative for next pixel row (scanline coherence):
1326                     if (_enableBlkFlagsHeuristics) {
1327                         // Use block flags if large pixel span and few crossings:
1328                         // ie mean(distance between crossings) is larger than
1329                         // 1 block size;
1330 
1331                         // fast check width:
1332                         maxX -= minX;
1333 
1334                         // if stroking: numCrossings /= 2
1335                         // => shift numCrossings by 1
1336                         // condition = (width / (numCrossings - 1)) > blockSize
1337                         useBlkFlags = (maxX > _BLK_SIZE) && (maxX >
1338                             (((numCrossings >> stroking) - 1) << _BLK_SIZE_LG));
1339 
1340                         if (DO_STATS) {
1341                             tmp = FloatMath.max(1,
1342                                     ((numCrossings >> stroking) - 1));
1343                             rdrCtx.stats.hist_tile_generator_encoding_dist
1344                                 .add(maxX / tmp);
1345                         }
1346                     }
1347                 } else {
1348                     _cache.clearAARow(lastY);
1349                 }
1350                 minX = _MAX_VALUE;
1351                 maxX = _MIN_VALUE;
1352             }
1353         } // scan line iterator
1354 
1355         // Emit final row
1356         y--;
1357         y >>= _SUBPIXEL_LG_POSITIONS_Y;
1358 
1359         // convert subpixel to pixel coordinate within boundaries:
1360         minX = FloatMath.max(minX, bboxx0) >> _SUBPIXEL_LG_POSITIONS_X;
1361         maxX = FloatMath.min(maxX, bboxx1) >> _SUBPIXEL_LG_POSITIONS_X;
1362 
1363         if (maxX >= minX) {
1364             // note: alpha array will be zeroed by copyAARow()
1365             // +1 because alpha [pix_minX; pix_maxX[
1366             // fix range [x0; x1[
1367             // note: if x1=bboxx1, then alpha is written up to bboxx1+1
1368             // inclusive: alpha[bboxx1] ignored then cleared and
1369             // alpha[bboxx1+1] == 0 (normally so never cleared after)
1370             copyAARow(_alpha, y, minX, maxX + 1, useBlkFlags);
1371         } else if (y != lastY) {
1372             _cache.clearAARow(y);
1373         }
1374 
1375         // update member:
1376         edgeCount = numCrossings;
1377         prevUseBlkFlags = useBlkFlags;
1378 
1379         if (DO_STATS) {
1380             // update max used mark
1381             activeEdgeMaxUsed = _arrayMaxUsed;
1382         }
1383     }
1384 
1385     boolean endRendering() {
1386         if (DO_MONITORS) {
1387             rdrCtx.stats.mon_rdr_endRendering.start();
1388         }
1389         if (edgeMinY == Integer.MAX_VALUE) {
1390             return false; // undefined edges bounds
1391         }
1392 
1393         // bounds as half-open intervals
1394         final int spminX = FloatMath.max(FloatMath.ceil_int(edgeMinX - 0.5f), boundsMinX);
1395         final int spmaxX = FloatMath.min(FloatMath.ceil_int(edgeMaxX - 0.5f), boundsMaxX);
1396 
1397         // edge Min/Max Y are already rounded to subpixels within bounds:
1398         final int spminY = edgeMinY;
1399         final int spmaxY = edgeMaxY;
1400 
1401         buckets_minY = spminY - boundsMinY;
1402         buckets_maxY = spmaxY - boundsMinY;
1403 
1404         if (DO_LOG_BOUNDS) {
1405             MarlinUtils.logInfo("edgesXY = [" + edgeMinX + " ... " + edgeMaxX
1406                                 + "[ [" + edgeMinY + " ... " + edgeMaxY + "[");
1407             MarlinUtils.logInfo("spXY    = [" + spminX + " ... " + spmaxX
1408                                 + "[ [" + spminY + " ... " + spmaxY + "[");
1409         }
1410 
1411         // test clipping for shapes out of bounds
1412         if ((spminX >= spmaxX) || (spminY >= spmaxY)) {
1413             return false;
1414         }
1415 
1416         // half open intervals
1417         // inclusive:
1418         final int pminX =  spminX                    >> SUBPIXEL_LG_POSITIONS_X;
1419         // exclusive:
1420         final int pmaxX = (spmaxX + SUBPIXEL_MASK_X) >> SUBPIXEL_LG_POSITIONS_X;
1421         // inclusive:
1422         final int pminY =  spminY                    >> SUBPIXEL_LG_POSITIONS_Y;
1423         // exclusive:
1424         final int pmaxY = (spmaxY + SUBPIXEL_MASK_Y) >> SUBPIXEL_LG_POSITIONS_Y;
1425 
1426         // store BBox to answer ptg.getBBox():
1427         this.cache.init(pminX, pminY, pmaxX, pmaxY);
1428 
1429         // Heuristics for using block flags:
1430         if (ENABLE_BLOCK_FLAGS) {
1431             enableBlkFlags = this.cache.useRLE;
1432             prevUseBlkFlags = enableBlkFlags && !ENABLE_BLOCK_FLAGS_HEURISTICS;
1433 
1434             if (enableBlkFlags) {
1435                 // ensure blockFlags array is large enough:
1436                 // note: +2 to ensure enough space left at end
1437                 final int blkLen = ((pmaxX - pminX) >> BLOCK_SIZE_LG) + 2;
1438                 if (blkLen > INITIAL_ARRAY) {
1439                     blkFlags = blkFlags_ref.getArray(blkLen);
1440                 }
1441             }
1442         }
1443 
1444         // memorize the rendering bounding box:
1445         /* note: bbox_spminX and bbox_spmaxX must be pixel boundaries
1446            to have correct coverage computation */
1447         // inclusive:
1448         bbox_spminX = pminX << SUBPIXEL_LG_POSITIONS_X;
1449         // exclusive:
1450         bbox_spmaxX = pmaxX << SUBPIXEL_LG_POSITIONS_X;
1451         // inclusive:
1452         bbox_spminY = spminY;
1453         // exclusive:
1454         bbox_spmaxY = spmaxY;
1455 
1456         if (DO_LOG_BOUNDS) {
1457             MarlinUtils.logInfo("pXY       = [" + pminX + " ... " + pmaxX
1458                                 + "[ [" + pminY + " ... " + pmaxY + "[");
1459             MarlinUtils.logInfo("bbox_spXY = [" + bbox_spminX + " ... "
1460                                 + bbox_spmaxX + "[ [" + bbox_spminY + " ... "
1461                                 + bbox_spmaxY + "[");
1462         }
1463 
1464         // Prepare alpha line:
1465         // add 2 to better deal with the last pixel in a pixel row.
1466         final int width = (pmaxX - pminX) + 2;
1467 
1468         // Useful when processing tile line by tile line
1469         if (width > INITIAL_AA_ARRAY) {
1470             if (DO_STATS) {
1471                 rdrCtx.stats.stat_array_renderer_alphaline.add(width);
1472             }
1473             alphaLine = alphaLine_ref.getArray(width);
1474         }
1475 
1476         // process first tile line:
1477         endRendering(pminY);
1478 
1479         return true;
1480     }
1481 
1482     private int bbox_spminX, bbox_spmaxX, bbox_spminY, bbox_spmaxY;
1483 
1484     void endRendering(final int pminY) {
1485         if (DO_MONITORS) {
1486             rdrCtx.stats.mon_rdr_endRendering_Y.start();
1487         }
1488 
1489         final int spminY       = pminY << SUBPIXEL_LG_POSITIONS_Y;
1490         final int fixed_spminY = FloatMath.max(bbox_spminY, spminY);
1491 
1492         // avoid rendering for last call to nextTile()
1493         if (fixed_spminY < bbox_spmaxY) {
1494             // process a complete tile line ie scanlines for 32 rows
1495             final int spmaxY = FloatMath.min(bbox_spmaxY, spminY + SUBPIXEL_TILE);
1496 
1497             // process tile line [0 - 32]
1498             cache.resetTileLine(pminY);
1499 
1500             // Process only one tile line:
1501             _endRendering(fixed_spminY, spmaxY);
1502         }
1503         if (DO_MONITORS) {
1504             rdrCtx.stats.mon_rdr_endRendering_Y.stop();
1505         }
1506     }
1507 
1508     void copyAARow(final int[] alphaRow,
1509                    final int pix_y, final int pix_from, final int pix_to,
1510                    final boolean useBlockFlags)
1511     {
1512         if (DO_MONITORS) {
1513             rdrCtx.stats.mon_rdr_copyAARow.start();
1514         }
1515         if (useBlockFlags) {
1516             if (DO_STATS) {
1517                 rdrCtx.stats.hist_tile_generator_encoding.add(1);
1518             }
1519             cache.copyAARowRLE_WithBlockFlags(blkFlags, alphaRow, pix_y, pix_from, pix_to);
1520         } else {
1521             if (DO_STATS) {
1522                 rdrCtx.stats.hist_tile_generator_encoding.add(0);
1523             }
1524             cache.copyAARowNoRLE(alphaRow, pix_y, pix_from, pix_to);
1525         }
1526         if (DO_MONITORS) {
1527             rdrCtx.stats.mon_rdr_copyAARow.stop();
1528         }
1529     }
1530 }